Deep learning and electrocardiography: systematic review of current techniques in cardiovascular disease diagnosis and management

Abstract

This paper reviews the recent advancements in the application of deep learning combined with electrocardiography (ECG) within the domain of cardiovascular diseases, systematically examining 198 high-quality publications. Through meticulous categorization and hierarchical segmentation, it provides an exhaustive depiction of the current landscape across various cardiovascular ailments. Our study aspires to furnish interested readers with a comprehensive guide, thereby igniting enthusiasm for further, in-depth exploration and research in this realm.

Introduction

Electrocardiogram (ECG) have been widely employed across various clinical disciplines. ECG represents one of the earliest non-invasive diagnostic methods in medicine, exerting profound influence on the field of cardiology [1]. However, accurate ECG interpretation typically necessitates substantial physician expertise, thereby fostering the gradual emergence of automated diagnostic technologies. In the 1970s, machine learning (ML) was introduced for automated ECG analysis, employing algorithmic modeling, image learning, and feature extraction to facilitate automated diagnosis [2]. Nevertheless, this approach was constrained by predefined rules and reliance on human-defined patterns, resulting in inadequate differentiation of subtle ECG differences and a relatively high rate of misdiagnosis. Consequently, ECG automated diagnostic programs have attracted considerable attention.

Over the past several decades, the rapid advancement of computational technology has catalyzed substantial growth at the intersection of computer science and biomedical research, generating numerous novel opportunities in healthcare diagnostics, therapy, and prognosis. Artificial Intelligence (AI), epitomizing the "Fourth Industrial Revolution", emulates human cognitive processes to harness sensor data for automated diagnostic solutions, enabling the identification of diverse abnormal patterns for diagnosing various diseases [3, 4]. Deep Learning (DL), a branch of AI, utilizes data-driven modeling, autonomous data recognition and learning, to generate powerful models, particularly suitable for handling heterogeneous data and dynamically changing ECG diagnoses [5].

The purpose of this review is to systematically survey the existing DL methods in conjunction with ECG and cardiovascular diseases, establishing a clear understanding of research hotspots. This review aims to clarify the current state of research concerning DL, ECG, and cardiovascular diseases, elucidate the future trends of DL and ECG applications in clinical settings, facilitate the advancement of research on DL, ECG, and clinical diseases, and expedite their translation into clinical practice for patient benefit.

Deep learning and the convergence with ECG

In the field of DL, Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have emerged as pivotal architectures. CNNs, initially conceptualized by LeCun et al. [6, 7] in 1989, excel in image feature extraction due to their local connectivity and weight sharing properties, making them ideal for ECG feature extraction [8, 9]. Their evolution has led to advanced models like AlexNet [10], ZFNet [11], VGGNet [12], GoogLeNet [13], and ResNet [14], enhancing DL performance. Conversely, RNNs, introduced by Rumelhart et al. [15] in 1986, are tailored for sequential data handling, with their ability to maintain context information via feedback loops. However, to tackle the vanishing gradient problem when dealing with long-term dependencies, LSTM [16]and GRU [17] were developed. The Convolutional Recurrent Neural Network (CRNN), a hybrid model born in 2015 [18], combines CNNs' spatial feature extraction prowess with RNNs' temporal dependency management, proving especially beneficial in ECG signal analysis. Meanwhile, Autoencoders (AEs) [19], another DL component, learn efficient data representations through compression and reconstruction, offering a powerful tool for dimensionality reduction and noise removal. Generative Adversarial Networks (GANs) [20], introduced by Goodfellow et al. in 2014, stand out for their capability to capture data distributions and generate new samples. They have been successfully applied to ECG data augmentation and denoising, overcoming the challenges posed by imbalanced or noisy data [21, 22]. This overview encapsulates the vibrant progression of DL, highlighting how researchers continuously innovate and refine these architectures to address limitations and broaden their applicability across various domains. This groundwork serves to offer readers a comprehensive insight into DL before delving into its specific applications in cardiovascular diseases.

The ECG data analyzed in our review were sourced from various channels, including: (1) exports from bedside ECG machines or cardiac monitors [23]; (2) data collected via ambulatory ECG devices [24]; (3) segments retrieved from public or specialized databases [25]; and (4) data derived from wearable devices [26]. These conventional ECG segments necessitate further processing before use. The data are stored in formats like SCP-ECG, DICOM, HL7 aECG, GDF, and more, which are designed to retain ECG waveforms, patient data, acquisition parameters, and diagnostic measurements. The ECG data processing workflow typically involves three stages: signal extraction and preprocessing, feature extraction, and classification. Given the low frequency and amplitude of ECG signals, coupled with interference from electromyographic activity, power line frequency, and baseline wander, preprocessing is essential to ensure accurate feature extraction and optimal model classification performance. Preprocessing techniques encompass both traditional and modern approaches. Traditional techniques include the use of FIR and IIR digital filters, while modern methods involve wavelet transform and adaptive filtering. Fourier and wavelet transforms are prevalent preprocessing tools, with the wavelet transform often preferred for its variable window length, providing superior resolution in both time and frequency domains, and greater accuracy than the Fourier transform [25]. Deep learning's role in ECG preprocessing is twofold: enhancing the model's data extraction capabilities through efficient filtering, and improving noise robustness through data augmentation strategies, such as the introduction of noise. The selection of these techniques hinges on the specific application, model architecture, and dataset size. Feature extraction strategies are categorized into hand-crafted methods, such as wavelet transforms, SVMs, kernel independent component analysis, and principal component analysis, and feature learning-based deep learning methods. The latter have significantly improved ECG signal classification accuracy and efficiency through automatic feature extraction, end-to-end models, attention mechanisms, management of multi-label imbalances, and deep transfer learning. ECG signal classification algorithms are divided into morphological methods, which include template matching and structural description, and methods based on signal features that emphasize waveform differences. The former is prone to noise interference and has a high computational cost and low accuracy, while the latter offers higher accuracy but may suffer from overfitting due to a lack of detailed heartbeat annotations. Manual classification and feature extraction methods, while valuable, have limitations in terms of accuracy and computational complexity for classifying ECG abnormalities and are being supplanted by DNN-based feature learning methods. DNNs enhance classification accuracy through autonomous feature extraction and minimize human intervention. CNNs, with their strength in processing one-dimensional temporal data, often serve as the basis for ECG signal analysis algorithms, enabling the construction of deep algorithmic models that effectively classify abnormal ECG [27]. The advent of deep learning models has undoubtedly advanced the accuracy of ECG signal classification and reduced reliance on human intervention, thus facilitating the integration of ECG with computational technologies and promoting the development and clinical adoption of innovative technologies.

Materials and methods

A systematic review was conducted to identify and aggregate published original studies that reported on DLlearning analyses of ECGs for the assessment of cardiovascular disease. The manuscript was prepared in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) guidelines [28]. This study utilized the Web of Science (WoS) platform, a multifaceted citation index database consisting of SCI, SSCI, A&HCI, CCR, and IC. The search strategy employed the following keywords: (‘Deep learning’) AND (‘Electrocardiography’ or ‘ECG’ or ‘EKG’) AND (‘cardiovascular disease’ or ‘heart disease’) with a publication date range from January 1, 2017, to September 30, 2023. This search yielded 1123 articles, which were systematically screened to exclude meeting abstracts (n = 221), proceeding papers (n = 134), non-English papers (n = 150), irrelevant articles (n = 180), duplicate articles (n = 33), no open access (n = 197) and poor quality (n = 10) articles, resulting in the selection of 198 high-quality articles post-discussion and voting by the research team. The data extraction in this process was carried out independently by multiple reviewers, with the aim of minimizing bias and errors as much as possible. The screening process is illustrated in Fig. 1. In this study, to elucidate the evolving trends and directions of deep learning applications in electrocardiogram (ECG) research, we conducted a statistical analysis of the publication years of the included articles (Fig. 2), the number of diseases investigated (Fig. 3), and the databases utilized (Fig. 4). For the graphical representation of our findings, we employed GraphPad Prism version 10.1.2 software. For bibliometric analysis and visualization of the selected literature, VOSviewer version 1.6.19 was used to map co-occurrences among countries/regions (Fig. 5) and keywords (Fig. 6). In Fig. 5, each dot denotes a distinct country, with its size reflecting the number of publications from that country. The lines connecting the nodes represent collaborative relationships between countries, where the thickness of the lines indicates the intensity of cooperation—thicker lines signify stronger collaborative ties. Figure 6 follows the same convention. Additionally, we summarized and ranked the top ten countries and regions based on article count (AC), citation count (CC), average citations per article (ACP), and total link strength (TLS) for the included articles (Table 1). We also summarized and ranked the top ten countries and regions based on the occurrences of keywords and their total link strength (TLS) (Table 2). To manage the wealth of information, the studies were classified according to disease types and key details were extracted, including article titles, publication years, source journals, datasets employed, sensors involved, and the specific deep learning application methods (Tables 3, 4, 5, 6, 7, 8, 9 and 10). Given the reliance solely on published data, ethical approval was unnecessary for this study.

Fig. 1

Screening flowchart

Fig. 2

Publication years of the included articles. The statistical data of the articles included in this study, analyzed across different years, demonstrate a consistent annual increase in trend

Fig. 3

Disease types in reviewed articles. The study's statistical analysis by disease type revealed a predominance of arrhythmia research, with less focus on valvular and myocardial diseases

Fig. 4

Types of databases used in the study. Database types commonly used for different diseases: public databases prevalent in arrhythmia research, while CAD/ACS and cardiac insufficiency/HF utilize more proprietary databases

Fig. 5

Principal countries and regions in the study. The top three in article count are China (59), USA (33), and Taiwan, China (30), while the citation count leaders are USA (1237), China (1071), and Singapore (728). In terms of average citations per country, the top three are Malaysia (158.75), Singapore (121.33), and Denmark (71.67). For international collaboration, the top three are China (29), England (24), and USA (23)

Fig. 6

Keywords of studies. The top three occurrences are deep learning (96), electrocardiogram (58), and classification (50), while the top three in Total link strength (TLS) are deep learning (415), electrocardiogram (249), and classification (215)

Table 1 Ranking of major countries and regions in the study

Table 2 Ranking of key research keywords

Table 3 CAD and ACS

Table 4 Cardiac insufficiency/HF

Table 5 Valvular heart disease

Table 6 Cardiomyopathy

Table 7 Arrhythmology

Table 8 Malignant arrhythmia and survival prognosis

Table 9 Blood pressure

Table 10 Others

Results

In this study, the 198 articles discussing the application of DL to ECG for cardiovascular disease (CVD) diagnosis were chronologically organized, reflecting a notable increase in annual publications starting from 2018 (Fig. 2). This surge could be linked to the advancement of AI technology, widespread adoption of smart devices, and intensified interdisciplinary medical research. The articles were categorized based on eight CVD types, with arrhythmology, blood pressure, and CAD and ACS accounting for the lion's share due to their prevalence, readily accessible public datasets, and the utility of DL with wearable devices (Fig. 3). In Fig. 4, public databases, particularly ARRDB[29, 30], AFDB[25, 31], and competition datasets[26, 27, 32, 33], played a crucial role in model development, especially for arrhythmology, blood pressure, and CAD/ACS, whereas proprietary databases were more prevalent in certain disease types like HF, cardiomyopathy, and valvular disease. Using VOSviewer, the global contributions were mapped, revealing China, USA, and China Taiwan as leading producers by article count, while USA topped the list for citation count (Fig. 5 and Table 1). Keyword analysis highlighted 'deep learning', 'electrocardiogram', and 'classification' as the most frequently occurring and interconnected terms (Fig. 6 and Table 2). Hybrid databases had less utilization. To evaluate the effectiveness of DL models, various performance metrics were employed, including area under the receiver operating characteristic curve (AUROC), accuracy (Acc), sensitivity (Sen), specificity (Spe), positive predictive value (PPV), negative predictive value (NPV), recall (Rec), precision (Pre), and false positives. This systematic approach elucidates the evolving landscape of DL applications in ECG-CVD research, identifies critical research areas, and underscores the importance of international collaboration and public dataset availability.

CAD and ACS

Studies based on PTB-ECG DB

The Physikalisch Technische Bundesanstalt Electrocardiogram Database (PTB-ECG DB) [34] is a large-scale database containing ECG recordings from subjects of varying ages and genders, encompassing health and diverse disease conditions. With its high sampling rate and precise resolution, this database serves as a foundation for numerous CAD and ACS studies.

Diagnosis of acute myocardial infarction (AMI) is crucial in clinical settings, prompting substantial research efforts in AMI detection. In 2017, Acharya et al. [35] achieved high accuracy using a CNN model on both denoised and noisy ECGs. Liu et al. [36] diagnosed extensive anterior wall MI with an Acc of 96.00% via multi-lead ECG. The model of various investigators[37–39] has excellent performance in MI diagnosis. Regarding MI localization, He et al.’s [40] MB-DenseNet-STSM model automatically labels valuable unlabelled samples, achieving an Acc of 96.09% in MI localization. The models established by several research groups [41–47] achieve nearly 100% accuracy in both diagnosis and localization, showcasing exceptional performance.

Studies based on other databases

Cho et al. [48]developed a DL model using Korean ECG data, achieving an AUROC of 0.902 in AMI diagnosis. Numerous scholars[49–55]continue to promote model innovation to diagnose MI. It is noteworthy that Liu et al. [54] proposed the AI-S strategy, reducing median ECG-to-catheterization-lab activation time from 6.0 min to 4.0 min (p < 0.01) and median door-to-balloon time from 69 to 61 min (p = 0.037); this is worth encouraging due to its clinical benefit. The collaborative research from many researchers [56,57,58] demonstrates remarkable proficiency in pinpointing the location of myocardial infarction and in the precise localization of the culprit vessels. Jin et al. [59] and Chaudhari et al. [60] constructed models assessing myocardial injury, with AUROC exceeding 0.760. Liu et al. [61] EfficientNet model accurately classified T-wave changes, ST-T segment changes, and pathological Q-waves. Tadesse et al. [62] proposed an end-to-end DL model incorporating transfer learning, demonstrating good performance in distinguishing MI at different stages. Han et al. [63] combined clinical doctors' diagnostic logic with a DenseNet network, establishing a model with an Acc and F1 of 93.65% and 94.27% in MI staging. The models constructed by many researchers [64,65,66,67,68] exhibits outstanding capability in discriminating CAD, obstructive coronary artery disease (obCAD) and evaluating coronary ischemia. Bhattacharya et al. [69] developed a dual-branch DL model integrating ECG and electronic medical records, showing good performance in predicting all-cause mortality and HF/stroke incidence. Eem et al. [70] assessed coronary artery calcium scores using a CNN model, achieving AUC and Acc of 0.890 and 80.6%, respectively. AMI and CAD are among the most common conditions in cardiology, with ECG serving as the fundamental diagnostic tool for these diseases. It is hoped that more research will be dedicated to their study, eventually enabling DL to diagnose MI and CAD efficiently, promptly, and accurately (refer to Table 3 for details).

Cardiac insufficiency/HF

Diagnosis of cardiac insufficiency often relies on laboratory tests such as BNP and echocardiography. However, ECG can now also serve as an auxiliary diagnostic tool. In 2018, Li et al. [71] used a CNN-RNN model to identify HF stages with an AUC and Acc of 0.851 and 97.6%, respectively. Vaid et al. [72] and Chen et al. [73] have developed models capable of accurately detecting different stages of HF and predicting adverse cardiovascular events. Sbrollini et al. [74], Chen et al. [4], and Botros et al. [75] have proposed models that demonstrate commendable performance in the diagnosis of heart HF. Raghu et al. [76] HFNet model, incorporating 12-lead ECG, age, and gender, identified HF patients with mPCWP > 18 mmHg with an AUROC of 0.81. The models conceived by Acharya [77]and fellow researchers [78,79,80] demonstrate noteworthy effectiveness in diagnosing congestive heart failure and forecasting episodes of acute decompensated heart failure. Attia et al. [81,82,83] CNN model performed excellently in diagnosing asymptomatic left ventricular dysfunction (ALVD) and left ventricular systolic dysfunction (LVSD), with AUCs of 0.93 and 0.82, respectively. Multiple independent investigators [84,85,86,87,88,89] have likewise augmented the precision in diagnosing LVSD by employing DL algorithms in their studies. In addition, Bachtiger et al. [90] CNN model diagnosed HFrEF through a specialized stethoscope with an AUROC and F1 of 0.85 and 0.369, respectively. Chen et al. [91] CNN model effectively recognized increased left ventricular end-diastolic diameter and predicted future cardiovascular risk. Liu et al. [92] CNN model judged abnormal BNP levels, with a model diagnostic AUC of 0.8934 for BNP ≥ 1000 pg/mL. It is expected that more research in the future will facilitate rapid and efficient diagnosis of cardiac insufficiency (refer to Table 4 for details).

Valvular heart disease

Echocardiography has traditionally been the primary means of detecting valvular heart disease. However, researchers now use ECG to diagnose these conditions as well. In 2020, Kwon et al. [93, 94] employed a CNN model to detect moderate-to-severe mitral regurgitation (MR) and developed an MLP-CNN combination model to diagnose moderate-to-severe aortic stenosis (AS), both achieving good AUCs and accuracies in internal and external validation sets. Elias et al. [95] employed ValveNet model accurately detected moderate-to-severe AS, MR, and aortic regurgitation (AR) through 12-lead ECG, with comparable diagnostic performance in white and black patients. Sawano et al. [96] employed 2D-CNN and FC-DNN combination model diagnosed moderate-to-severe AR with an AUROC and Acc of 0.802 and 82.3%, respectively. Additionally, Vaid et al. [97] employed MLP-CNN model achieved ideal results in diagnosing moderate-to-severe MR and AS, with AUROCs approaching 0.9. These studies further substantiate the potential of deep learning in diagnosing valvular heart diseases (refer to Table 5 for details).

Cardiomyopathy

While echocardiography and cardiac MRI have traditionally been the mainstay for diagnosing cardiomyopathies, researchers can now effectively diagnose these conditions using ECGs. Kokubo et al.’s [98] ENN model accurately diagnosed left ventricular dilatation (LVD) and left ventricular hypertrophy (LVH) through 12-lead ECG, with AUROCs of 0.810 and 0.784, respectively. Following continuous endeavors by investigators [99,100,101,102,103], the AUROC for the diagnosis of LVH using the developed model ascended to a notable peak of 0.98, indicating exceptional discriminatory capacity. Hypertrophic cardiomyopathy (HCM) is a commonly encountered cardiomyopathy in clinical settings. Ko et al. [104] used a CNN model, leveraging a large amount of HCM data from the Mayo Clinic digital [105], to efficiently diagnose HCM, achieving an AUC and Acc of 0.96. Maanja et al. [106] and Siontis et al. [107] have lowered the false positive rate in their models' diagnosis of HCM to 13.5%, while concurrently increasing the AUC to 0.98. Chen et al. [108] employed the VGG-16 model for genotyping patients with HCM, achieving AUC of 0.89 for diagnosing G + individuals, which outperforms the Toronto score [109] (AUC = 0.69). Lee et al. [110] focused on peripartum cardiomyopathy (PPCM) and achieved promising results. Mutations in the Phospholamban (PLN) gene can lead to arrhythmogenic cardiomyopathy (ACM) and dilated cardiomyopathy, among other diseases [111]. Bleijendaal et al. [112], together with Lopes et al. [113], and others, have developed various DL models for diagnosing PLN p. Arg14del-related cardiomyopathy, with the best-performing model achieving an AUC of 0.96, demonstrating overall superiority over expert cardiologists (AUC = 0.91). Gumpfer et al. [114] combined a CNN algorithm with clinical information to develop a composite model for diagnosing myocardial scar (MS), with the model diagnosing MS with an AUC and Acc of 0.89 and 78.0%, respectively. These studies not only enrich the technical arsenal for diagnosing cardiomyopathies, but also provide strong support for clinical decision-making (refer to Table 6 for details).

Arrhythmology

Based on MIT-BIH database

The MIT-BIH database, encompasses over 4,000 long-term ambulatory ECG recordings. Comprising databases such as AFDB, ARRDB, and NSRDB, it includes data on various arrhythmias like NSR, AF, SVT, CAVB, AVB, PVC, providing valuable resources for related research. Andersen et al. [115] utilized a CNN-RNN combination model to diagnose AF based on RR interval segments, achieving an AUC and Acc of 0.947 and 87.40%, Wu et al. [116] achieved improvements up to 0.9983 and 97.56% with Morlet-CWT-CNN model. Thereafter, the models developed by numerous researchers [25, 31, 117,118,119,120,121,122,123,124] have demonstrated substantial improvements in the diagnostic performance for AF, with accuracies approaching nearly 100%. Dang et al. [125] introduced the MSF-CNN B model, efficiently diagnosing multiple arrhythmias through single-lead ECGs, with mean Acc, Sen, and Spe of 98.00%, 96.17%, and 96.38%, respectively. Successive multiple studies [30, 126,127,128,129] have incrementally improved the classification performance across various types of cardiac rhythms, with accuracies nearing 100%, accompanied by a marked increase in computational speed. It is worth mentioning that Zhang et al. [130] designed a BAM-ResNet model that achieved a Pre of 98.95% in categorizing atrioventricular nodal reentry tachycardia (AVNRT) and atrioventricular reentrant tachycardia (AVRT). Marco et al. [131] introduced a MobileNetv2 model, which displayed outstanding diagnostic performance in an imbalanced dataset for premature ventricular contractions (PVCs), Acc of 0.9990 and AUC of 0.9963. Additionally, CNN model [132] and KNN model [29] by Yu et al., as well as the CNN model from Sarshar et al. [133], and the ResNet-18 model developed by Ullah et al. [134], have all demonstrated commendable performance.

Based on competition databases

Competition databases refer to those officially released for public competitions, offering ECG and other data resources to interdisciplinary researchers, such as Computing in Cardiology Challenge and CPSC 2018. A plethora of studies are founded on these databases. Pourbabaee et al. [135] employed a CNN model to diagnose PAF, achieving high correct classification rates and low false-negative rates on the PAF prediction challenge database. Fan et al. [136] proposed the MS-CNN model, utilizing different filter sizes in a dual-stream convolutional network to capture multi-scale features, diagnosing AF with an AUC and Pre of 98.13% and 91.78%. Subsequently, researchers have developed several models, including the LSTM-RNN model [137], DCAA model [138], ResNet-BLSTM/RBF model [32] DDCNN model [33], a ResNet-based model[139], and CNN-based models [140,141,142,143], all of which have contributed to enhancing the diagnostic performance for AF. It should be mentioned that Zhang et al.’s [140] GH-MS-CNN model, incorporating hybrid multi-scale convolution modules, improved AF diagnostic performance, with mean Acc, Pre, and F1 of 0.9984, 0.9989, and 0.9954, respectively. To enhance model performance and ensure their robustness across various databases, numerous researchers have undertaken additional work: Zhang et al.’s [144] MCNN-BLSTM model dynamically set branches and enhanced feature information, Xiong et al.’s [26] 37-layer CRN model utilizing a transfer generator to expand sample size, Hu et al.’s [145] model based on residual blocks and transformers, and Li et al.’s [146] SC-CNN model employing a self-complementary attention mechanism, all showed superior performance in AF diagnosis. Tutuko et al.’s [147] CNNs-BiLSTM model recognized AF on QTDB and LUDB datasets [148] with Acc, Pre, and F1 of 99.79%, 99.01%, and 98.96%, respectively. In addition, Dhyani et al. [149] and Ganeshkumar et al. [27] each proposed a model classifying nine types of arrhythmia signals, with best Acc and Rec of 0.962 and 0.949.

Based on other databases

Lai et al. [150] trained a CNN model on Holter data from a Chinese hospital, diagnosing AF through II-lead data with Acc, Sen, and Spe of 93.1%, 93.1%, and 93.4%, respectively. CNN-based models [23, 151,152,153,154,155,156], DDNN-based models[157], and the CAT model proposed by Yang et al. [158] have all demonstrated commendable performance in the diagnosis of AF, with reported AUCs and Acc for detecting AF occurrence reaching up to 99.4% and 97.9%. It is worth noting that the Confident learning-CNN model proposed by Chen et al. [23] achieved an AUC of 0.935 for atrial fibrillation (AF) recognition on the KGHDB dataset [159].

Moreover, several researchers have conducted extensive studies[24, 160,161,162,163,164] and achieved commendable outcomes in the diagnosis of various cardiac arrhythmias, localization of PVC, and identification of Concealed Accessory Pathway (AP). Such as Wang et al.’s [160] Densenet 169 (R) model identified AP with AUC and Acc of 0.941 and 0.973. DL models have also proven instrumental in the localization of the origin sites of cardiac arrhythmias. Zhang et al.'s [165] ECGNet model has demonstrated appreciable accuracy in pinpointing the origin of PVC. Similarly, studies by Missel et al. [166], Monaci et al. [167], and Pilia et al. [168] have showcased commendable localization precision for the origins of ventricular tachycardia (VT), evidenced by robust performance in both internal and external validation tests, such as CNN-VAE model localization errors for LV pacing sites averaging just 5.3 ± 2.6 mm.

DL models have impressively demonstrated their mettle in predicting arrhythmia onset [169,170,171], AF risk stratification [172], and postoperative recurrence [173,174,175], thus substantiating the edge of deep learning. Notably, the CNN-CatBoost model by Tang et al. [173] excels in forecasting AF relapse with an AUC of 0.859, surpassing traditional AF ablation prognosis measures like the APPLE Score [176] (AUC = 0.644) and the CHA2DS2-VaSC index [177] (AUC = 0.650). Together, these studies poignantly highlight the expansive applicability of DL in arrhythmia prognosis and assessment across diverse clinical contexts (refer to Table 7 for details).

Malignant arrhythmia and survival prognosis

Life-threatening arrhythmias are disorders that can swiftly disrupt cardiovascular dynamics, potentially leading to hemodynamic collapse, syncope, and even sudden death upon manifestation. Hence, early detection plays a pivotal role. Multiple research teams have developed DL models [178,179,180,181] that achieve strikingly high accuracies of close to 99% in the identification of VT and VF events. Elola et al. [182] employed Bayesian-optimized DNN models to discriminate between pulseless electrical activity (PEA) and pulse-generating rhythm (PR). TSENG et al. [183] developed a CNN model capable of issuing warnings 50 and 20 s before ventricular fibrillation (VF) onset, achieving accuracies of 56% and 47.1%, respectively. Kaspal's team [184] reported a CNN-RCN model with a remarkable 93.24% accuracy in predicting sudden cardiac deaths (SCD). Raghunath et al. [185] utilized a DNN model to predict one-year all-cause mortality from ECGs, with an AUC of 0.876, outperforming the fragrance risk score (AUC = 0.648) [186] and the Charlson comorbidity index (AUC = 0.816) [187]. Capretz et al. [188] adopted a CNN model integrating ECG and baseline data to forecast 30-day acute myocardial infarction (AMI) and mortality risks in chest pain patients, achieving AUROC, Ppv, and Npv of 93.9%, 73.6%, and 99.9%, respectively. Tsai et al. [189], Kondo et al. [190] with their CNN model, and Sun et al. [191] with their ResNet-based model, all demonstrated high predictive precision in forecasting short-term or long-term outcomes for patients. Prifti et al. [192] and Doldi et al. [193] have developed DL models with remarkable diagnostic precision in identifying Long QT Syndrome (LQTS), exhibiting high accuracy. The KanResWide model by Diaw et al. [194] demonstrates a mean overall absolute error (MGAE) in assessing QT intervals of merely 11.2 ± 12.1 ms. Liu et al.’s [195] proposed CNN-BiLSTM model and Liao et al.’s [196] ResNet-18 model exhibit superior predictive accuracy in diagnosing Brugada type 1 syndrome, achieving an AUC value approximating 0.97. Dunn et al. [197, 198] and ElRefai et al. [199] harness DL models to screen out non-appropriate populations for S-ICD, utilizing the T:R ratio to predict TWOS risk with a mean absolute error (MAE) of only 0.0461. Nejadeh et al. [200] and Wouters et al. [201] forecast CRT outcomes with models demonstrating enhanced predictive performance (C-statistic = 0.69), surpassing QRSAREA [202] (C-statistic = 0.61). Collectively, these studies illuminate the potential and advantages of deep learning in recognizing and anticipating malignant arrhythmias (refer to Table 8 for details).

Blood pressure

Model performance standards in this section reference the Association for the Advancement of Medical Instrumentation (AAMI) and British Hypertension Society (BHS). AAMI requires MD of BP values < 5 mmHg and STD < 8 mmHg in cohorts with over 85 participants. BHS categorizes accuracy into A, B, and C grades based on absolute error: < 5 mmHg proportion > 60%/50%/40%, < 10 mmHg proportion > 85%/75%/65%, < 15 mmHg proportion > 95%/90%/80%, corresponding to A, B, and C grades, respectively. These criteria assess the accuracy and reliability of BP measurement models.

Soh et al. [203] deployed a CNN model to diagnose hypertension via ECG, utilizing data from MIT-BIH NSRDB and the SHAREE database [204]. Their model demonstrated outstanding performance, with Acc, Sen, Spe, and Ppv nearing 100%. Miao et al. [205] introduced a ResLSTM model that evaluated blood pressure using single-lead ECGs, meeting the American Association for Medical Instrumentation (AAMI) standards, and achieving A-grade performance according to the British Hypertension Society (BHS) criteria for MAP and DBP measurements. Multiple DL models [206,207,208,209,210,211,212,213,214] assessed blood pressure through ECG and PPG signals, conforming to both BHS and AAMI standards. For instance, BPNet model proposed by Long et al. [206] integrated preprocessing, cross-modal fusion, post-feature extraction, and multitasking modules, constructing the model using CNN and Feature Pyramid Network (FPN) fusion of PPG and ECG signal features. Yang et al. [215] presented a hybrid CNN + LSTM + Dense model combining ECG, PPG, and physiological data to evaluate blood pressure, satisfying AAMI standards and achieving A and B grades according to BHS. Furthermore, DL models have been applied to predict intraoperative hypotension (IOH), defined as MAP < 65 mmHg during surgery. Jo et al. [216] devised a ResNet model that, using ECG and arterial blood pressure (ABP) data, predicted IOH 3 and 5 min prior with AUROCs of 0.970 and 0.935, and AUPRCs of 0.943 and 0.882, respectively (refer to Table 9 for details).

Others

Fan et al. [217] put forth LSTM and BiLSTM models capable of forecasting elderly health status over the following day using single-lead ECG, achieving impressive AUROC and Accuracy rates of 0.9312 and 93.21% for health status prediction, respectively. Butt et al. [218] introduced the SGDM-AlexNet-CNN model, which classifies falls and activities with a Validation Accuracy of 98.44%. Attia et al. [219] developed a CNN-based model that utilizes 12-lead ECG to determine gender with a Sex Prediction AUC of 0.97 and Accuracy of 90.4%. Several researchers [172, 220,221,222,223] constructed DL models to estimate ECG-derived age, achieving a best mean absolute error (MAE) of 5.8 ± 3.9 years; they also found that individuals with a significant discrepancy between ECG-age and actual age face a significantly elevated risk of death [220, 223]. Iakunchykova et al. [224] utilized a CNN model revealing a statistically significant correlation (r = 0.12, p < 0.0001) between heart delta age (HDA) and brain delta age (BDA). Mori et al. [225] introduced a CNN-LSTM model for diagnosing autism spectrum disorder (ASD) in minors, achieving AUC and Accuracy values of 0.95 and 0.89, respectively. Lou et al. [226] proposed a CNN model to diagnose Left Atrium Enlargement (LAE), reporting AUCs of 0.8688 and 0.8990 for moderate and severe LAE, along with C-indices of 0.688 and 0.806 for predicting Stroke and AF occurrences. Liu et al. [227] combined D-dimer, ECG, and Chest X-ray (CXR) to diagnose aortic dissection (AD) with an AUC of 0.943. Another study by Liu et al. [228] presented a DL model to diagnose Acute Pericarditis using 12-lead ECGs, attaining an AUC, Sensitivity, and Specificity of 0.954, 78.9%, and 97.7%; when coupled with the STEMI-DLM [53], specificity improved to 99.4%. Gomes et al. [229] developed a DL model proficient at classifying arrhythmic ECGs and ECGs from COVID-19 patients. Hassan et al. [230] demonstrated excellent performance of their DL model in diagnosing COVID-19 in patients with heart disease. Lastly, Gupta et al. [231] constructed a DCM model capable of identifying fetal arrhythmia (FA) through Fetal ECG, achieving an AUC, accuracy, and precision of 99.57%, 94.18%, and 95.63% for FA diagnosis (refer to Table 10 for details).

Discussion

Through a comprehensive review of existing research, the application of DL in ECG analysis has become increasingly common, with CNN and RNN serving as the core frameworks and continuously undergoing innovative improvements. However, ECG signals are characterized by high noise and complexity, posing stringent requirements for preprocessing techniques. Factors such as baseline wander and electrode position can affect signal quality and model performance. Therefore, optimizing preprocessing techniques, including noise reduction and standardization, has become a crucial breakthrough. From the aforementioned research, it is evident that datasets, as the core of DL model construction, exhibit two main characteristics in their selection and application. Firstly, in the early stages of research, scholars tend to adopt publicly available datasets, such as the MIT-BIH database, which are easily accessible and comprehensive, greatly simplifying the data collation process and providing convenience for interdisciplinary and junior researchers. Secondly, there is a focus on the application of public and proprietary datasets in disease model construction. Public datasets are commonly used for model training in diseases such as coronary artery disease (CAD) and arrhythmias, while proprietary datasets are more often seen in conditions like heart failure (HF) and cardiomyopathy. These differences stem from factors such as the difficulty of data collection, the purpose of database creation, and the openness of public datasets. Nevertheless, the inherent differences among various datasets may significantly affect model performance, necessitating a clear understanding of the specificities of each dataset. It is noteworthy that transfer learning has demonstrated significant value in addressing issues of data scarcity and heterogeneity, aiding in reducing training time and enhancing model effectiveness. However, its complex parameter tuning and limited flexibility limit its application prospects, necessitating further exploration.

Multiple studies reviewed here are based on local lead ECGs from wearable devices, particularly single-lead ECGs, which show great potential in daily health management. Nevertheless, wearable devices also face numerous challenges. For instance, static ECG models struggle to adapt to dynamic conditions, and noise and artifacts can lead to false alarms. While single-lead data offer convenience, it may also result in the loss of critical information. Additionally, the issue of protecting the privacy of health data has become prominent. Currently, most models exhibit significant limitations when faced with the complex and variable ECG diagnoses encountered in clinical practice. There is an urgent need to develop a comprehensive model with strong generalization ability and high recognition accuracy. This necessitates the establishment of large-scale, multi-center databases covering diverse disease types and geographical regions, combined with continuous optimization of model performance through advanced model structures and feedback mechanisms.

It is common knowledge that the rigor, comprehensiveness, and feasibility of experimental design directly affect the reliability of research outcomes. In early DL and ECG studies, experimental designs were relatively simplistic, and while dividing datasets for training and testing yielded good performance indicators, they did not fully demonstrate the practical application value of the models. Nowadays, experimental designs have become more sophisticated, introducing validation sets, external test sets, and multi-dimensional evaluation metrics to comprehensively assess model performance. Methods such as tenfold cross-validation are also employed to reduce data selection biases. However, no matter how optimized, experimental designs may still conceal potential issues. For example, limited data sources may lead to selection bias, and the complexity of clinical diseases can also affect ECG interpretation. Therefore, when evaluating model performance, it is essential to consider not only numerical indicators but also the prerequisites and clinical application significance of model construction. Additionally, the standardization of electrocardiographic data represents a critical issue for future research, as there is currently a lack of standardized ECG input types and preprocessing protocols. As previously mentioned, the ECG data utilized in studies are often derived from different devices, and preprocessing protocols vary across different researches. It is important to emphasize that model performance may vary due to differences in signal acquisition and processing methods. Therefore, we cannot solely judge the performance of a model based on numerical differences; similarly, it is uncertain which form or preprocessing method can fully unleash the potential of deep learning techniques. Moreover, the heterogeneity of ECG signals demands models with robust generalization capabilities to maintain consistent performance across diverse patient populations and clinical settings. Models also require continuous updates and iterations to retain their clinical relevance. Thus, the application of deep learning in ECG requires not only continuous exploration and discovery of new model characteristics by each researcher, but also the establishment of unified standards within the industry to provide a solid platform.

Currently, some studies have successfully integrated model performance with clinical applications, such as the AI-S active alert strategy [54] in reducing diagnosis time for MI patients and successful cases in screening non-adaptive populations for S-ICD [198] and predicting VF onset [183]. These examples demonstrate the practical utility of models in clinical practice. However, questions remain regarding how to ensure the rationality of model performance improvements, their alignment with clinical logic, and their specific clinical value. It is recognized that many deep learning models are termed "black boxes", owing not only to their complex internal structures, intractable decision paths, and the uncertainty inherent in their training processes, but also to the insufficient exploration of disease-relevant information embedded within the signals during their construction. This diminishes the models' efficacy and interpretability, thereby impeding their application in clinical practice. Clinical practitioners value the reliability of evidence-based medicine; without understanding how a model arrives at specific diagnoses, there is a lack of trust in the model's decision-making process. Even though some models have been compared to the diagnostic acumen of clinical physicians, showing a slight superiority in overall diagnostic performance [86, 185], this indicates that significant improvements are needed before deep learning models can be applied in clinical settings. Additionally, the acceptance of novel assistive diagnostic devices by patients and their families is a concern, especially in remote or medically underserved hospitals. Furthermore, ECG analysis models, as medical devices, must navigate stringent regulatory approval processes, such as FDA's 510(k) clearance and the European CE marking. These approval processes can be both time-consuming and costly, limiting the models' ability to quickly enter clinical practice. Moreover, medical devices must undergo rigorous testing and evaluation to ensure their safety and efficacy. In this process, the number of algorithms may serve as an indicator of a product's maturity and diversity. Thus, the assessment of the number of algorithms in medical devices is also a crucial step. Therefore, despite the rapid development of deep learning, there is no evidence to suggest that the role of human experts in ECG analysis will be eliminated. It should be emphasized that deep learning algorithms are designed for computer-assisted interpretation, serving as a decision support system to assist human experts, rather than replacing them. In the foreseeable future, the central role of trained cardiologists in ECG analysis remains unassailable.

In the future, we look forward to more studies that can balance the feasibility of clinical applications with the robustness of model performance, jointly advancing the field of DL and ECG to provide more accurate and efficient diagnosis and treatment solutions for patients.

Limitations

In this study, the latest research on the application of deep learning in electrocardiography has been analyzed from multiple perspectives to reveal the hot spots in this field. However, for the construction of some deep learning models, such as the process of computer development programs, design concepts, and methods, only a general outline of development has been shown, without intuitive presentation of the details, and more attention has been paid to their relevance to clinical applications. In addition, this study selected the Web of Science literature search platform under the premise of complying with the (PRISMA) guidelines, and if databases such as PubMed or Scopus are added for search, it may result in a larger sample size and reduce selection bias.

Conclusion

This article focuses on the application of deep learning and ECG in the field of cardiovascular diseases. It systematically reviews the research trajectory encompassing 198 articles, presenting them in a chronological order to closely align with clinical research practices. The application prospects of AI in the medical field are vast, but to build reliable DL models that meet clinical needs requires not only the support of a vast database but also close collaboration among clinicians, data researchers, and programmers. Additionally, we must recognize that while AI is powerful, it cannot replace the professional knowledge of doctors. Instead, it should serve as an auxiliary tool to enhance diagnostic accuracy, treatment precision, and prognostic improvement capabilities. Otherwise, when significant clinical errors occur in AI, issues of responsibility attribution become inevitable. Furthermore, the integration of AI with clinical practice may affect the doctor–patient relationship, which requires careful consideration based on the development level of AI and policy formulation. These issues are complex and challenging, necessitating our collective efforts to address and resolve them.